letta-ai/letta-code

What it solves

Letta Code provides a stateful agent harness that allows AI agents to move beyond simple tools and behave more like people. It addresses the limitation of short-term memory in LLMs by giving agents a persistent identity, a sense of experience over time, and the the ability to evolve through self-improvement.

How it works

The system uses a combination of memory blocks and "dreaming" (periodic self-reflection) to allow agents to programmatically rewrite their own context, system prompts, and skills. It utilizes MemFS, a git-tracked context system, to sync agent state across environments. Agents can be deployed across multiple remote environments (laptops, VMs, cloud sandboxes) and can be accessed via a CLI, desktop app, or messaging integrations like Slack and Telegram.

Who it’s for

It is designed for developers and users who want to create proactive, always-on agents that can learn from their interactions and maintain long-term state across different platforms.

Highlights

• Self-improvement: Agents can rewrite their own prompts and skills to adapt over time.
• Multi-agent capabilities: Agents can call other agents (or themselves) as subagents, either synchronously or asynchronously.
• MemFS: Git-based tracking of all agent context and memory blocks.
• Cross-platform access: Unified agent access via CLI, desktop app, browser, and various messaging channels.
• Remote execution: Ability to run agents across multiple remote environments and machines.